On Sat, 11 Mar 2017, Kyle Huey wrote: > static void init_intel_misc_features(struct cpuinfo_x86 *c) > { > u64 msr; > > + if (rdmsrl_safe(MSR_MISC_FEATURES_ENABLES, &msr)) > + return; > + > + msr = 0; > + wrmsrl(MSR_MISC_FEATURES_ENABLES, msr); > + this_cpu_write(msr_misc_features_shadow, msr); > + > if (!rdmsrl_safe(MSR_PLATFORM_INFO, &msr)) { > if (msr & MSR_PLATFORM_INFO_CPUID_FAULT) > set_cpu_cap(c, X86_FEATURE_CPUID_FAULT); > } > > probe_xeon_phi_r3mwait(c); The way you are doing it breaks the ring3 mwait feature because you overwrite the R3MWAIT bit the first time you update the MSR on context switch. > } What you really want is to fixup the r3mwait part: static void probe_xeon_phi_r3mwait(c) { .... if (ring3mwait_disabled) return; set_cpu_cap(c, X86_FEATURE_RING3MWAIT); this_cpu_or(msr_misc_features_shadow, 1UL << MSR_MISC_FEATURE_ENABLES_RING3MWAIT_BIT); if (c == &boot_cpu_data) ELF_HWCAP2 |= HWCAP2_RING3MWAIT; } and have a function for cpuid fault: static void init_cpuid_fault(c) { if (rdmsrl_safe(MSR_PLATFORM_INFO, &msr)) return; if ((msr & MSR_PLATFORM_INFO_CPUID_FAULT)) set_cpu_cap(c, X86_FEATURE_CPUID_FAULT); } and then do: static void init_intel_misc_features(struct cpuinfo_x86 *c) { u64 msr; if (rdmsrl_safe(MSR_MISC_FEATURES_ENABLES, &msr)) return; /* Clear all MISC features */ this_cpu_write(msr_misc_features_shadow, 0); /* Check the features and update the shadow control bits */ init_cpuid_fault(c); probe_xeon_phi_r3mwait(c); msr_write(MSR_MISC_FEATURE_ENABLES, this_cpu_read(msr_misc_features_shadow); } Hmm? Thanks, tglx